[PATCH] NOMMU: Improved handling of get_unmapped_area() errors

The attached patch does two things:

 (1) We no longer check the return value of file->f_op->get_unmapped_area()
     unless we actually called it. We know addr is zero otherwise because
     we'd've given an error earlier if it wasn't.

 (2) If -ENOSYS was returned by that operation, then we assume we actually
     called a driver (such as the framebuffer driver) that might want to
     invoke the operation in a lower level driver (such as matroxfb) if one
     exists, and that it found that one didn't.

     We translate the -ENOSYS error into -ENODEV - the error we would have
     given if the operation was not supplied in the file ops.

     Doing this permits us an opportunity for arch_get_unmapped_area() or
     something else to be called if we want that to happen, particularly in
     the MMU case.
Signed-Off-By: David Howells <dhowells@redhat.com>
Signed-Off-By: Linus Torvalds <torvalds@osdl.org>

mm/nommu.c

@@ -567,13 +567,15 @@ unsigned long do_mmap_pgoff(struct file *file,
 	 * that it represents a valid section of the address space
 	 * - this is the hook for quasi-memory character devices
 	 */
-	if (file && file->f_op->get_unmapped_area)
+	if (file && file->f_op->get_unmapped_area) {
 		addr = file->f_op->get_unmapped_area(file, addr, len, pgoff, flags);
-
 		if (IS_ERR((void *) addr)) {
 			ret = addr;
+			if (ret == (unsigned long) -ENOSYS)
+				ret = (unsigned long) -ENODEV;
 			goto error;
 		}
+	}
 
 	/* we're going to need a VMA struct as well */
 	vma = kmalloc(sizeof(struct vm_area_struct), GFP_KERNEL);